What actually happened: Washington took a drastic step

On Friday, June 13, Anthropic — currently the most valuable AI startup in the world with a valuation nearing a trillion dollars — received an order from the U.S. government unlike anything in the history of the industry. The company had to immediately deactivate its two most advanced models for all foreign customers, including some of its own employees who are not U.S. citizens.

The reason? According to Washington, someone obtained information about a method to bypass the security restrictions of the Fable 5 model — a so-called jailbreak. Anthropic claims it was a narrow attack revealing several previously known and less serious vulnerabilities that other publicly available models can also find. The government, however, decided to act immediately and without further details.

The entire incident is all the more intriguing because it unfolds amid an open dispute between Anthropic and the Donald Trump administration. Defense Secretary Pete Hegseth had previously publicly labeled the company a "supply chain risk" — the first time this designation had been applied to an American company. Anthropic sued the Pentagon over this and the court ruled in its favor.

What the Fable 5 and Mythos 5 models can actually do

To understand the scope of the problem, we need to look at what these models are truly capable of. Adam Hanka, Director of Data and AI at Creative Dock and co-host of the podcast "Carbon vs. Silicon," described Mythos's abilities very specifically for Forbes:

"Mythos was purposefully trained to read and analyze large amounts of code and work autonomously for extended periods. Thanks to this, it can operate for several days without human intervention," Hanka explains. The model creates its own hypothesis about a vulnerability, tests it by running a test instance in a server environment, and after verification moves on to the next one. "The average user imagines progress in models as better answers in ChatGPT. In this case, it's a complex agentic workflow that looks like calling that chat a thousand times in a row."

Filip Dřímalka, founder of Aibility and author of the book "The Future of Non-Work," confirms that models like Fable 5 are, according to independent rankings (such as Artificial Analysis), still unsurpassed. "Some open-weight models are catching up, however. But it's not simple — running these models is hardware-intensive and the investments can be quite high, while delivering lower performance."

Hanka adds that Asian models — DeepSeek, Qwen, Kimi, or GLM — are rapidly closing in on the Western elite, but they tend to reach the level of GPT or Gemini, not the level of Mythos. "There is a possibility, however, that somewhere in China there are laboratories with more advanced models that are not available to ordinary users."

Why the ban is a warning shot for Europe

According to Hanka, Washington has crossed an imaginary line. "The United States government intervened in which model will be available to whom. It thereby confirmed fears that dependence on non-European technology is an enormous security risk for us and a major vulnerability."

Dřímalka points to a structural problem: "The development of top models has long ceased to depend solely on capital and talent. It's increasingly about using the best models to develop the next generations, and catching up there is extremely difficult. Add to that hardware — data centers and energy — where similar geopolitical restrictions are already being addressed today."

Both experts agree that instead of building its own capacity, Europe spent too long erecting barriers. "Europe must invest more in technology, science, and research, that's clear. But the current situation is also caused by the fact that Europe simply lacks a common capital market where sufficient capital can be gathered for private entities," says Hanka.

Is Fable 5 really as dangerous as the government claims?

The answer is not black and white. Hanka points out that no safeguards are one hundred percent effective — but that applies to all current AI models, not just Fable. "Anthropic states that the model underwent thousands of hours of testing by teams from the U.S. government, the UK government's AI Safety Institute, and others. None of them managed to find a way to broadly bypass the model's protections."

Dřímalka is more cautious: "There are many cases where people have managed to bypass the guardrails of large companies. Every company will tell you its products are safe. But if there's a company on the market that approaches AI responsibly, in my view it's Anthropic. I would trust the company more than the Donald Trump administration."

Also interesting is the argument that the global suspension of access to the best models deprived cybersecurity defenders in particular of a crucial tool. Hanka calls it a double-edged sword: "The attacker uses the same tool to search for vulnerabilities and the defender to patch them. But if there are many vulnerabilities, a stronger tool gives more power to the attackers. They only need to find one vulnerability."

Dřímalka adds a practical dimension: "When you shut off access to the models, the defender who has built it into their processes loses the tool — while attackers often keep alternatives."

What Europe — and Czechia — can do

The situation is not hopeless. Both experts offer concrete paths that Europe should take.

Dřímalka emphasizes that local and open-weight models can eventually catch up to the top tier. "As is already evident today, similar quality can be achieved by cleverly combining multiple models. That's also why we, as Europe, should strive to gain at least partial technological independence."

Europe is not starting from zero. French Mistral AI is among the world's leading producers of large language models, and its models are available within European cloud infrastructure as well. The Czech Republic has also recently launched the Czech AI Factory project in Ostrava, which aims to become one of the nodes of European artificial intelligence. For Czech companies, it is crucial that open-weight models like Meta's Llama or DeepSeek can be run on their own hardware without dependence on American clouds.

Hanka warns, however, that without a common European capital market, it will be difficult to raise the kind of funds that American companies have managed to secure. Anthropic alone has raised tens of billions of dollars in investments over the past year — sums that European AI startups can only dream of.

"In our region, we have a strong foundation and talents, and we should build on that. It's not just about models, but primarily about their availability and control. From a security perspective, using top-tier models is no longer an optional add-on — it's a fixed part of risk management with full accountability of company management," Dřímalka concludes.